Nonlinear dynamics are being widely used nowadays in neuroscience to characterise complex systems. Thus, hidden potential of the dynamical properties of the physiological phenomenon can be detected by these approaches especially to elucidate the complex human brain activity gathered from the electroencephalographic (EEG) signals. A reliable and non-invasive measurement of memory load, to measure continuously while performing a cognitive task, is highly desirable to prevent decision-making errors. Such measurements help to avoid cognitive overload, especially at high mental or physical workload places. 

We analyzed the publicly available physionet EEG database to develop the machine learning framework. From the EEG recordings of 36 subjects, linear and non linear dynamic features are extracted. First the features are balanced between the two groups using SMOTE algorithm, then using the random forest algorithm, the significant features are selected. Further, the ALASSO algorithm is used on the top 10 features selected by random forest to yield Higuchi's Fractal dimension, Theta power spectral density and Standard Deviation of the signals as the features having a significant discriminatory power. Running algorithm of Decision Tree(DT), this model was trained, and cross-validated, to develop a performance prediction model, the SMORASO-DT (SMOTE+RANDOM FOREST+ALASSO-DECISION TREE).

The SMORASO identified three highly predictive variables for performance stratification. The SVM Fine Gaussian kernel and Decision Tree predicted performance with 73% and 78% accuracy respectively. We identified three predictors, Higuchi's Fractal dimension, Theta power spectral density and Standard Deviation of the signal that were highly predictive of classification accuracy and mental working memory load complexity.

Subject stratification using machine learning methods will help to identify people who can tolerate high cognitive load. Additionally, we identified new biomarkers for mental arithmetic task performance stratification.